When AI memory clings to what is no longer true

When an old priority returns

Imagine a company changing its strategy. For two years it focused on large accounts. Now it is concentrating on small and medium-sized businesses. The decision has been made, explained and documented.

A few weeks later, its AI assistant prepares a prospecting plan. It recommends large accounts. It has retrieved presentations, meeting minutes and sales objectives that all point towards that priority. They are numerous, precise and entirely authentic.

And outdated.

The assistant invented nothing. It used information that had ceased to carry authority. In this example, memory becomes a brake on change.

Accurate information can lead to a poor decision

We expect a great deal from AI memory: retrieving our documents, remembering our preferences and saving us from repeating what we have already explained.

But a preference can change. An instruction can be withdrawn. A decision can be replaced. Information long considered certain can become contested.

The question “Does the AI remember it?” leaves another question open: “Does it still know under which circumstances it may use it?”

What was true yesterday belongs to our history. What can guide action today requires further examination.

Learning to measure correct forgetting

Md Nayem Uddin and his co-authors address precisely this challenge in From Recall to Forgetting. Their Memora benchmark presents agents with simulated conversations in which memories are added, changed and deleted. Their FAMA metric penalises the reuse of information that has become invalid.

In their analysis of 25 recommendation errors from the best-performing agent for that task, 16 came from outdated information that had not been forgotten. This finding concerns a small sample of errors, not all responses or all AI systems. The conversations are synthetic and evaluation relies heavily on automated judges. Nevertheless, it highlights a concrete failure that simple recall tests can miss.

I see this as a requirement to add to our criteria for trust: checking that memory can stop using what has been explicitly invalidated.

Preserve a trace, withdraw its authority

Should we therefore delete old decisions? We would often lose what helps us understand a journey.

In a company, an abandoned strategy can explain an investment or an organisational structure. In a family history, corrected testimony can illuminate how a narrative developed. In personal notes, a discarded project can remain an important step.

The distinction I want to explore for Mnemory is this: a trace can retain its historical value while losing its authority over the present.

Our company’s former strategy could help answer “Why did we recruit that team?” It should no longer determine the answer to “Whom should we approach tomorrow?”

This distinction requires knowing the scope and provenance of information, and the decisions that changed its status. It also requires respecting deletion requests: systematically keeping an archive cannot be the default response.

Who decides what is no longer true?

The latest message does not automatically become the truth.

A remark may be a hypothesis. A contradiction may reflect two different situations. Two people may have divergent memories without the machine being able to settle the matter.

An AI might flag: “This new instruction seems to contradict the previous one. Does it replace it?” The answer should belong to the person authorised to make that decision.

For Mnemory Knowledge, this opens a discussion about the authority of decisions and procedures. For Legacy, about the coexistence of testimonies and corrections. For Notes, about our ability to develop our projects without continually being pulled back to old intentions.

These are design principles to examine and test. They are not an announcement of features already available.

Memory that can accompany change

The test could start simply: give an assistant an instruction, explicitly replace it, then present several situations in which the old instruction would once have been relevant.

Does it apply the new decision? Does it ask for clarification when the scope remains uncertain? Does it reintroduce the old instruction from a summary that has not been updated?

These behaviours need to be measured together. An assistant that stops answering altogether might avoid outdated memories without having useful memory. A few successful tests would not guarantee the absence of all residual influence either.

What interests me is memory we can rely on when we change our minds, our direction or our understanding.

Trustworthy memory must let us evolve without trapping us in who we used to be.